Consensus and Volatility in Presidential Approval
by
Paul Gronke
Department of Political Science
Duke University
Durham, NC 27708-0204
email: gronke@acpub.duke.edu
October 31, 1996
Abstract
One of the central themes in research, teaching, and popular commentary about American politics is the power of the Presidency. This is no surprise, since American domestic politics and
policymaking, and even many decisions on the world stage, rotate about the axis of the U.S. president. Presidential approval plays a particularly critical role in presidential power, as ample studies
have demonstrated over the past twenty years. We must be concerned, then, over recent evidence of
an increasingly fickle American public. Presidents have enjoyed new peaks but also suffered new
lows in popular approval. Changes in approval seem to be more rapid and the diversity of opinions
at any one time seems to be growing. If the American public has become more volatile and less consensual in their evaluations of the president, it will be difficult for presidents to build and maintain
legislative coalitions, unify domestic actors behind foreign policy initiatives, and ultimately maintain their electoral coalitions. The president’s success in swaying public opinion and the president’s
ability to use public support as a leadership tool may depend as much on the amount of agreement
(consensus) and the lability of public support (volatility) as it does on the average level of approval.
This paper takes an opening look at measuring and modeling volatility and consensus in presidential
approval. Drawing on data from the New York Times/CBS polls from 1976–1993, the paper shows,
via descriptive analysis, how mean approval, over time volatility, and cross sectional variability, are
distinct processes. The paper then proposes some alternative ways to examine volatility and consensus.
Paper prepared for the annual meeting of the Southern Political Science Association,
Atlanta, GA, November 1996. The author would like to thank John Aldrich, John Brehm,
Lyle Scruggs, Renée Smith, Janet Box–Steffensmeier, and Jim Stimson for
encouragement and help, and the ICPSR (and Suzanne DeBoef) for access to the NYT
data. All remaining errors are certainly the fault of the author.
Draft Copy: Comments and criticism highly encouraged
Introduction
Presidential approval is a key political resource in an era of the public relations presidency.
When approval ratings are higher, members of the president’s party in Congress are less
likely to be defeated in the midterm and the president’s legislative proposals are more successful. The presidential agenda, both on the foreign and domestic side, is “bolder” under
conditions of high approval.1 And presidents enjoying high levels of approval are more
likely to be reelected. The president’s performance in this “perpetual election” (Simon
and Ostrom, 1989) or “new referendum” (Brace and Hinckley, 1992) is clearly a key to
understanding presidential power in the postwar era (Neustadt, 1990).
The conventional method for understanding changes in presidential approval has been
to model movements in aggregate mean approval as a function of time in office, economic
change, and “rally events.”2 Models of the mean, however, abandon an enormous quantity
of information, namely, the attitudes of individual respondents. One recent piece (Brehm
and Gronke, 1995) attempts to extract individual response variation via change in the variance portion of the time series. These results are briefly described below.
There is another way to “extract” individual level data: use the individual responses.
While the move from aggregate to individual level data may seem obvious, there are many
theoretical and technical difficulties associated with the change, many of which I have not
surmounted by the time of this essay. Here, I present some preliminary ideas about modeling consensus and volatility using combined cross sectional surveys and aggregate political
and economic data. I start by showing how individual and aggregate level variability are
different theoretically, but how they are empirically difficult to disentangle. Next, I present
some graphics which illustrate the basic differences which I hope to study. Finally, I present
some suggestions about analytic strategies for these data, providing what I hope is sufficient
grist for discussion and criticism.
Defining Volatility and Consensus
The presidential approval series is quite familiar to most political scientists. As shown
in Figure One, presidential approval has meaningful ebbs and flows, in response to the
1
These claims are well established in the literature on elections, presidential approval, and presidential
leadership. The citations are too numerous to include here. Four recent volumes are Brody (1991), Brace
and Hinckley (1992), Kernell (1993), and King and Ragsdale (1988). For evidence on legislative success, see
Rivers and Rose (1985) and for evidence on ”boldness” of agenda, defined as breadth of legislative proposals,
frequency of vetoes, and use of military force, see Simon and Ostrom (1989).
2
C.f. Jackman, 1995; MacKuen, Erikson, and Stimson, 1992; Simon and Ostrom, 1989; Ostrom and
Simon, 1985; Brody, 1991; Brace and Hinckley, 1992; Kernell, 1983, 1978; Mueller, 1973.
1
state of the economy, elections, and presidentially relevant events.3 Approval, plotted from
1976–1993, shows a relatively steady decline during Carter’s term, with a “rally” after the
Iran hostage crisis, followed by an even steeper fall; a honeymoon after Ronald Reagan’s
election, followed by a drop during the recession of 1981, a long climb until the Iran-Contra
hearings; another honeymoon after Bush’s reelection, a high point of popularity during the
Gulf war, then a free fall up to the election of Clinton. These patterns are well understood
by the political science profession (see citations listed above). However, I would like to take
another look at the approval series, and suggest that our understanding of what presidential
approval means has, until now, been missing a key element: the shape of the underlying
distribution.
Presidential approval is theoretically a continuous variable (see, for example, Ostrom
and Smith, 1992; Ostrom and Simon, 1988), yet we only observe dichotomous responses,
disapprove or approve.4 While I do not have a special insight into the original justification
for the item, its use through the years suggests that Gallup saw it as analogous to a “thermostatic” measure of public opinion about the president (see also MacKuen, Erikson, and
Stimson, 1995). Some suggest, in fact, that the aggregate measure is all that is of interest, and we should ignore individual ratings of presidential job performance (Brehm and
Gronke, 1994). Still, one could imagine a distribution of approval ratings, ranging from
those who have a visceral reaction against everything the President says and does, to those
who blindly endorse all presidential actions, where the bulk of Americans probably fall
somewhere in the middle. If one accepts this premise, then one might also ask whether the
shape of this theoretical distribution — peaked or flat, unimodal or multimodal — has any
implications for public opinion about the president.
3
The figure here plots aggregate presidential approval from the New York Times/CBS polls, conducted
on approximately a monthly basis, from 1976 through 1993 (from the ICPSR archive). The item used is
identical to that used by Gallup: “Do you approve or disapprove of the way (PRESIDENT) is handling his
job as president?” The figures here correlate at .99 with the Gallup data.
4
The National Election Studies has used a five point job approval scale, but only infrequently.
2
I think that it does, and the way to conceptualize changes in the distribution is via
volatility and consensus. By volatility, I mean the degree to which presidential approval
changes in response to roughly equivalent shocks across administrations. This is an aggregate, over time statistic. An administration where presidential approval decays more
under one kind of shock than another administration under the same kind of shock reflects
a more volatile public. Approval ratings that are likely to slip sharply in one direction or
another are “volatile” ratings. Alternatively, volatility in approval can be conceptualized as
the relative weight of the short term components of presidential approval (e.g. more rapid
or steeper growth in the “coalition of minorities” (Mueller, 1973) or greater responsiveness
to economic change), holding events constant.5
5
From this perspective, events are “shocks” that we need to control for in order to gain leverage on the
relative weight of the short and long term components.
3
By consensus, I mean the spread of the distribution of presidential approval at any
single time point. A less consensual public reveals a relatively wide variance in approval,
regardless of the mean. Widespread agreement on the performance of the president, on the
other hand, implies a tight variance about the cross–sectional mean. Finally, consensus can
be examined over time, in what I will call consensus–building or dissensus–causing shifts
in approval. An example of consensus building events were the series of Soviet actions after
World War II, which Page and Shapiro cite as a cause of the very high levels of agreement
about the proper American position toward the Soviets (Page and Shapiro, 1992, Ch. 5).
Examples of dissensus–building events were the U.S. actions in Vietnam in the early 1970s,
which Zaller (1992) shows resulted in wide partisan disagreement on the U.S. role. This
project moves beyond past studies by considering the shape as well as the central tendency
of the distribution of presidential approval.
Should we care about volatility and consensus? Increases in volatility and decreases
in consensus could hurt the president’s ability to build support in Congress and pass his
legislative program. From the perspective of a rational political actor deciding whether
to support or oppose a president, the possibility for rapid changes in approval introduces
uncertainty into decisions. If a president’s job approval is relatively stable, Congress, the
media, and the public have a consistent image of the President. Political actors can make
calculations without bothering to include uncertainty. A president with a wildly variable
base of support, in contrast, introduces a host of uncertainties into the political system.
Should I tie my fortunes to a popular President today, when he may be unpopular tomorrow? This dovetails with claims that modern presidents are highly sensitive to changes in
opinion polls (Brace and Hinckley, 1992). Presidents delay certain actions because they
cannot afford a drop in the polls or pursue other activities, especially foreign policy adventures, because they expect a boost in popularity. George Bush received criticism for
attempting to protect his 1991 post Gulf War popularity surge into the 1992 election by
purposely adopting a limited domestic agenda. What this research can help us answer is
whether recent presidents have had to become more sensitive to public opinion because the
public has become increasingly volatile. If change is important, the amount of change is as
well.
Volatility allows us to test the hypothesis that presidential approval has become fundamentally less stable over time, with a resulting decline in the president’s ability to act. Like
many summary political judgements, presidential approval has both a long term and a short
term component. The long–term, stable component of presidential approval includes party
affiliation, strength of party attachments, and ideological beliefs (Mueller, 1973; Fiorina,
1981). Short term influences, those changes and events in the world which lead to pertur4
bations in approval, include economic performance, wars, scandals, and what have come
to be called “rally events” (Mueller, 1973; Kernell, 1978). If we conceive of volatility as
the relative weight of short term components on approval, holding events to the side, then
an increase in volatility implies that the president is no longer able to rely on a core base of
support during his term in office. Presidential approval, and presidential power along with
it, becomes a victim of events that are largely outside of the president’s control.
If we conceive of volatility more broadly as the degree to which approval changes in
response to comparable “shocks” and consensus as the level of intergroup agreement on
the president’s performance, then volatility and consensus will lead to a more nuanced
picture of presidential “rallies.” Extensive prior research, starting with Mueller (1973), has
demonstrated the importance of key political events to presidential approval. Some events,
such as wars, assassination attempts, and good economic news, boost the presidential stock
while other events, such as scandals and civil discontent, slash away at presidential support.
Conventionally, these events have been described narrowly as “rally points” and “negative
events.”6 Yet, rallies and busts contain implicit claims about consensus. In this respect,
presidential rallies can be viewed like the stock market. The daily stock report tells us
not only the rise or fall of the Dow (movement in the mean), but also whether gainers out
paced losers (the cross sectional variance). A rally across the board is different than a rise
marked by peaks and valleys — particularly if you’ve invested in the wrong stock! The
same applies to presidential approval, and to public opinion more generally. Rather than
leading only to a boost or drop in mean approval, we also want to know whether certain
kinds of events might be associated with changes in consensus.7
Volatility and consensus, in tandem, provide us with the tools to explore changing patterns of presidential leadership. If the Neustadtian president was able to marshall generalized public support as a regular tool of leadership, the Kernellian president has to constantly
appeal to a labile, fickle public (Neustadt, 1990; Kernell, 1993). If Kernell is right, and the
combined impact of the Vietnam War, civil discontent, and the new style of press coverage
led to a fundamental restructuring of the presidency, we ought to observe some evidence
of this in presidential approval. These kinds of changes will not be evident in the mean,
however, but in the new responsiveness of the public to events, i.e., volatility and consensus. Similarly, the different kind of political circumstances described in a historical study
6
For example, see Simon and Ostrom (1989), Brace and Hinckley (1992), and Kernell (1978). Brody
(1991) provides a nice critique of this approach.
7
For example, scholars of international relations and public opinion and foreign policy describe a fundamental shift in the public’s attitude toward U.S. involvement in foreign policy conflicts (e.g. Holsti, 1992;
Jentleson, 1992). The implication is that the post–Vietnam public ought to react differently — less consensually – to foreign policy engagements.
5
of presidential leadership (Skowronek, 1993) should be evident in the pattern of public
response during those presidencies. A president during the “politics of disjunction,” such
as Carter, is unable to establish credibility and is “suddenly vulnerable;” while a president
like Roosevelt or Reagan, who practices a “politics of reconstruction” is able to rebuild
the regime in a “new and more relevant way.” (Skowronek, 1993, Ch. 2). These descriptions sound, without doing too much damage to this seminal work, like one president who
encounters widespread dissensus, Carter, and two presidents who rebuilt consensus. My
approach provides a quantitative lens to these qualitative accounts of presidential style and
leadership.
Limits to Conventional Wisdom Regarding Volatility and Consensus
It remains to be seen whether the empirical representation of approval as a dichotomy
provides any leverage at all on the shape of this theoretically continuous distribution. At
the aggregate level, of course, approval is measured in a more continuous fashion, bounded
only as a proportion, at 0 and 1.0. This has led scholars to examine variability in approval
via the standard deviation. The evidence in this respect is spotty and inconclusive. One
set of analysts suggest that “instability has not increased steadily over time” (Gallup and
Edwards, 1990, p.122), while another set suggest that volatility increased during George
Bush’s term (Brace and Hinckley, 1992).
The technique of estimating the standard deviation of approval ratings over an administration compensates for unusual highs and lows, but does not recognize the turbulence of
history. Some administrations have a high frequency of events (e.g., the Johnson administration) while others have a relatively low frequency of events (e.g., the Carter administration), yet the standard deviation of the two administrations are relatively close. Does
the slightly higher standard deviation of approval under Johnson indicate a more “volatile”
public than that under Carter, even if fewer consequential events took place? Similarly,
how can one compare the standard deviation during Nixon’s term with any other without
accounting for Vietnam protests and the Watergate scandal? Comparing standard deviations alone is the wrong way to examine volatility. If the empirical referent for volatility is
the variance in the approval equation, then this should be modeled directly.
Consensus, in contrast, has seldom been directly featured in any studies of presidential approval. I do not mean to suggest that disagreement has not been recognized. The
“politics that presidents make,” historically, has been a politics of taking advantage of partisan divisions and repudiating rising opposition (Skowronek, 1993). Kernell (1993), Brace
and Hinckley (1992), and Brody (1991) all make reference to the level of public disagreement. But all the evidence is indirect, either newspaper accounts, measures of legislative
6
opposition, or measures of support among party groups in the population. This study, to
my knowledge, will be the first to examine subgroup variation within the cross sections
directly. As such, it can build a new conventional wisdom about public consensus.
Measurement Issues: Time Series, Cross Sections, and Pooled
Cointegrated Time Series
I have reviewed a number of theoretical reasons why volatility and consesus should enhance
our understanding of presidential approval. To summarize, our methodology should be
chosen so as to allow us to examine:
Spread in approval over time. Have movements in approval become larger or smaller,
more or less frequent?
Spread in approval at any time point
Predictability of approval. Has the stochastic element of approval increased?
Next, I discuss three different methods of examining approval, each of which corresponds
roughly to the research areas. The most important choice to be made is the appropriate
level of analysis. Should we work with the approval in the aggregate, as a set of integrated
time series, a set of pooled time series, or pooled cross sections? This is not a choice that
can be determined by methodological preference; these choices may allow us to answer
certain kinds of research questions and not others.
At the highest level of aggregation, volatility can be modeled in the stochastic porIn
a time series, volatility can be measured, and modeled, in the stochastic portion of a time
series. Measuring volatility and consensus in a cross section is substantially more problematic. While I can provide some suggestive ways to examine consensus, all involve
significant assumptions about the structure of presidential approval.
tion of a time series model. Measuring volatility and consensus in a cross section is
substantially more problematic, in large part due to the binomial nature of the dependent
variable. Finally, recent work on cointegrated time series holds some promise for answering
the question of causality — has the approval series become more of less predictable over
time. These techniques are quite difficult, however, and some controversy surrounds their
applicability.
7
Modeling Volatility in A Time Series8
Scholars of the presidency know that we cannot understand mean levels of presidential approval without knowing the particular circumstances of each administration, whether this
includes economic performance, time in office, and war casualties (Mueller, 1973), media
coverage of the administration (Brody, 1991), or the “dramatic events of a term” (Brace
and Hinckley, 1992, p. 10). Brehm and Gronke (1994, 1995) model the variance of presidential approval in a similar fashion. Typically, the analyst chooses an appropriate function
to reparameterize the mean level of approval () in terms of the substantive explanatory
variables, while leaving 2 as a constant:
= f (X; )
(1)
What has been long neglected in studies of presidential approval is the opportunity to repa-
rameterize 2 and to identify independent variables that help explain changes in the variance. Nor have there been any attempts to extend these models down to the individual level.
Neither of these steps are merely methodological exercises. The aim is explicitly substantive: it allows the analyst to test hypotheses that implicate both over time and individual
level indicators of political change on the president’s ability to muster popular support.
In previous work, we dervied an estimator which combines two classes of models of
heteroskedasticity in time-series (Brehm and Gronke, 1994, 1995). One approach to modeling variance in a time series (“multiplicative heteroskedasticity”) treats variance as a function of some set of explanatory variables, but is not conditional on variance in prior periods
(Greene, 1993, p. 405–407; Harvey, 1990; for application to presidential popularity, see
Beck, 1991). An alternative approach (“autoregressive conditional heteroskedasticity” or
ARCH) treats variance as conditional on variance in previous periods, yet typically omits
explanatory variables (Greene, 1993, p. 438-442; Harvey, 1990). We combined these two,
retaining the autoregressive nature of the time series is retained, but adding a series of
explanatory variables. The parameterization for the variance thus becomes:
t
= 0 + 1 2t,1 + Z
(2)
where 0 and 1 represent the ARCH parameters, Z is a matrix of explanatory variables,
and Z represents the impact of other explanatory variables on . This straightforward
extension of Engle and Beck’s work reveals a statistically significant increase in volatility
over the post war era. The level of volatility is responsive to real world events.
8
This material in this section is presented in more detail in Brehm and Gronke, 1995
8
Modeling Consensus and Volatility in a Time Series of Cross Sections
The cross sectional analogue of volatility is consensus: the degree of agreement or disagreement at any time point on presidential performance. What is substantially more problematic, however, is deriving an analogous measure for cross sections. Ideally, I would
operationalize consensus as the variance within any single cross–section. This is impossible, given the binomial nature of the approval variable. In the cross section, there is no
measure of the standard error that is not a simple transformation of the level of approval.
Lacking access to a continuous distribution, there is another way to gain leverage on consensus. The underlying construct that I wish to represent is the level of disagreement on
the president’s performance, both within a cross section, and during significant historical
events (thus comparing cross sections over time). An alternative way to think about consensus, and one which relates more directly to the problem of building and maintaining
support coalitions, is to identify the positions of major groups within the U.S. population,
and to think of consensus as a process of groups unifying and dividing on their level of
approval. A more heterogeneous set of approval ratings across groups indicates lower consensus, which a unified set of ratings indicates high levels of agreement.
One quantitative technique for measuring heterogeneity on some characteristic across a
set of groups has been developed elsewhere as a measure of population diversity (Gronke,
1993; see also Sullivan, 1973). Essentially, the statistic represents the probability that a
pair of randomly selected individuals from a population will both “agree” (have the same
value) on the characteristic under study, or, in our case, both approve of the president’s
performance:
A=1,(
X Y )=V
p
k =1
2
k
(3)
where A = the probability of agreement, Yk = the proportion of the population falling
in a given category within each of the groups (e.g. the marginals), V = the number of
variables (groups) which the population is broken down into, and p = the total number of
categories across all variables (in our case, 2 V ). The critical assumption underlying the
technique is that the process of categorizing individuals into groups completely represents
the underlying distribution of approval, thus each group member has an equal probability
of “approving.” If the groups are chosen carefully, this assumption does not do serious
damage to our inferences. A consensus time series could be created out of cross sectional
data, and be employed as a dependent variable in regression analyses.
The problem with this technique is that it treats all political cleavages equally, so that
diversity among religious, economic, or racial groups receive identical weights. Second,
9
among what categories do we calculate diversity? One possibility would be to calculate a
measure heterogeneity or homogeneity within the group who approved of the president’s
performance. The argument would be that a more heterogeneous set of approvers indicates
more consensual appeal. On the other hand, we may want to calculate diversity across the
groups, calculating the likelihood that one “approver” and one “disapprover” (our sample
pair) differs on the characteristics under study. My intuition is that these two ways of
calculating diversity are simple transformations of each other, but I have not yet attempted
to demonstrate this.
As a first cut at population diversity, I will instead rely on a more simple, straightforward measure of variance. I plot below the relative levels of agreement among politically
consequential groups in the population. For example, how do Blacks and whites or Republicans, Independents, and Democrats differ in their evaluations of presidential performance?
This procedure is no different than producing a set of time series on approval, aggregating from each randomly selected cross section, and pooling the time series. For each time
point, an overall measure can easily be derived using the standard variance measure for
pooled data, calculated across j groups, each with n respondents:
P (X , X ) n
V ariance = (P n ) , 1
2
j
t
j
j
This is a strictly descriptive measure, but may provide some insight into ebbs and flow in
variance of approval. When the measure is low, the average group response is close to the
grand mean, thus indicating a higher level of consensus.
Volatility may also be conceived of in another way, as the relative weight of short and
long term considerations used by citizens when they evaluate the president. If the public has
become more volatile over time, then regression estimates on, say, strength of partisanship
ought to decline over time, while the weights attached to economic change and influential
events should increase. Taken to a logical extreme, this representation implies a least square
dummy variable (LSDV) specification of the approval model:
Y
nt
= nt + t + n + Xnt k + nt
(4)
for n = (1:::N ) cross sections over t = (1:::T ) time points (Sayrs, 1989). Unit effects are
assumed for each time point. LSDV, however, does not speak to the question of whether
the ’s vary over time. Furthermore, pooled cross sectional procedures within at least two
statistical packages, SAS and STATA, assume panel data, e.g. time series on the same units
pooled over time. This means that I cannot use the individual level data, but may proceed
using the data aggregated into groups. While this strikes me as a serious loss of information,
it has been suggested that, so long as I aggregate upon variables that are contained in the
10
regression equation, going from cross sections to time series entails no loss of information.
I have put off this step for now.
Assessing Causality for Cointegrated Time Series
A third direction for assessing the impact of individual level opinions and demographic
characteristic as well as aggregate measures of important political events and economic
changes lies somewhat in between purely aggregate analyses and purely cross sectional
analyses. Three recent studies ask research questions like those proposed here (Ostrom
and Smith, 1992; Box–Steffensmeier and Smith, 1996; Box–Steffensmeier and DeBoef,
1996). Ostrom and Smith (1992) give a clear exposition of the cointegration technique,
with a comprehensive summary of how previous work on approval fits into this framework.
The problem with the work for the purposes outlined here, however, is that the equilibrium
model that they propose, if I understand them correctly, is premised on a steady state of
approval, perturbed by various events. I’m not sure that the model can allow for changes in
the volatility of approval, nor how cointegration tests for causality (essentially, whether sets
of time series move in tandem), relate to conventional notions of causality in the regression
framework. These uncertainties are probably due to my unfamiliarity with the approach.
In a similar vein, Box–Steffensmeier and Smith (1996) demonstrate that the macropartisanship series (MacKuen, Erikson, and Stimson, 1989) is fractionally integrated, meaning
that the “aggregate measures of Republican and Democratic identifiers exhibit neither stationarity nor unit root behavior.” This suggests both that there is behavioral heterogeneity in
response to political events, and that the microfoundations of macropartisanship are heterogeneous across the population, but also that shocks to macropartisanship are relatively long
lasting.9 Box–Steffensmeier and DeBoef (1996) broaden the scope of the analysis, comparing movements in macroideology to macropartisanship, but more importantly for my
purposes, move one level of analysis down, disaggregating the macropartisanship series
into levels of political sophistication. They find a stronger link between macropartisanship
and macroideology among more sophisticated portions of the public.
The question these studies raise, besides the necessity of learning new methodology,
is whether it is more productive to abandon attempts to model approval at the individual
level, while pooling the cross sections, and instead move to a middle level of aggregation,
and take advantage of these innovations. I do not claim to have even minimal capabilities
with these techniques, but they seem to be a promising area for future research. I comment
more on these options below.
9
The extent of responsiveness to system shocks in the macropartisanship series relates directly to changes
in presidential approval in response to events.
11
Suggestive Data and Future Directions
So far, I have presented a number of ideas about consensus and volatility, and listed a
series of alternative ways to examine them. I close by presenting some preliminary data on
variability in approval, mainly as suggestive directions for future work.
Do mean approval and variance in approval move in tandem? I examine this question
in two steps. First, I plot presidential approval, broken down by party ID and race. The
patterns, shown in Figures Two and Three, are not unexpected. The starkest differences in
approval are evident during the Reagan years, where the average level of approval among
Democrats was approximately 40%, versus nearly 90% among Republicans. Carter’s and
Bush’s partisan pattern of approval (distinct from their overall levels of approval, reported
in the first figure), are somewhat similar. Both experienced the expected pattern, with
inpartisans approving at higher levels than out partisans, yet the gap was seldom more than
30%. Similarly, Black/Non-Black differences were muted during the Carter years, as the
average rating among both groups declined during the late 70s. Blacks and non-blacks
gave substantially different ratings, on average, to Reagan, varying by as much as 50 points
during the height of the 1981 recession. The pattern during the Bush years is something of
a mix, neither as divided as during Reagan, nor as unified as during Carter.10
10
I apologize to the reader for not removing the erroneous data point just after 19May87. I do not believe
that Blacks were approving of Reagan’s job performance at an 80% rate, 20 points higher than non-blacks.
12
It is easy to jump ahead, and make an eyeball estimate of variability in these ratings.
Partisan differences on presidential performance expand and contract, and not always in
tandem with rises or declines in average approval. For example, honeymoons (rises in mean
approval) are clearly associated with consensus in approval. On the other hand, Reagan’s
approval rating during the 81-82 recession remained more stable among Republicans than
among other partisan groups, and Bush’s post Gulf War drop was much more severe among
Democrats and Independents. And the gaps between Black and white expand and contract
in dramatic fashion over this time period; at times (e.g. late in Carter’s term, early in
Bush’s and Clinton’s term), Blacks and whites were in agreement. During other periods,
the differences are as great as we usually associate with the American racial cleavage. All
this is not news, of course, but it reinforces the point that the spread of approval varies in
important and distinct ways from mean approval.
13
The next figure present these same data, but converted into an overall summary measure
of variance. The top frame reproduces the mean approval series for ease of comparison,
while the bottom frames shows the variance between partisan and racial groups. I will only
point out a few features in these plots. First, notice, as was already evident in the previous
plots, that Reagan’s presidency was marked by long period of high variance in approval,
regardless of the rises and falls in mean approval. Reagan’s presidency, from these data,
divided the public, on partisan and racial grounds, even though conventional interpretations
of his presidency is one in which he enjoyed relatively high levels of public acclaim (e.g
Brody, 1991, pg. 10–15). The uniqueness of the Reagan years in terms of trend – he
failed to display a linear decline – is also apparent in terms of variability. George Bush,
on the other hand, whose average level of popularity is described by Brace and Hinckley
as more volatile than most other presidents (1994, pg. 144), actually experienced rather
low variablity in ratings among partisan groups (not racial groups), with spikes and drops
associated with the Gulf War conflict and the onset of the economic decline prior to the
1992 election.
14
Second, rapid changes in approval (for example, increases during the three honeymoon
periods covered by this series, and decreases during the 1991-2 recession and six months
after the Gulf War) are associated with decreases in variability. This confirms findings
of Brehm and Gronke (1995), where we found the most dramatic decreases in volatility
duing and after a presidential election (1995, pg. 21). If these results hold up under further
scrutiny, they indicate that honeymoons are rallies in all respects, with increases in level
and consensus of approval. But why should elections lead to greater agreement? My
expectation is that periods of rapid changes in approval are also periods of widespread
uncertainty, where the public is exposed to conflicting interpretations of the presidential
performance. However, it is under conditions of dual information flows that the public
seems best able to reach informed, reasoned opinions (Zaller, 1992; Sniderman, Brody,
and Tetlock, 199x). Perhaps this is what is operating here.
Taking this one step further, elite debate is most often organized on partisan lines, and
this should be reflected in a great amount of variability by party. I have some indirect
evidence on this point. The variability about the mean for racial groups is about half of that
for partisan groups. On average, Blacks and whites varied 8.5 points (s.d. 0.4) from the
15
mean, whereas partisan groups varied 16.9 points (s.d. 0.5) from the mean. Excluding the
Reagan years does not reduce this gap, but controlling for party affiliation of whites (not
done here) will reduce it. A nice comparison for these data would be plots broken down,
again by party and race, during the 1960s, when elite messages were coded as often by race
as they were by party.
Future Directions
This paper takes an exploratory look at volatility and consensus in presidential approval.
I some steps toward extending previous work to the individual and pooled cross sectional
level. As a work in progress, I have no conclusions to present, only a series of questions.
The dynamics of presidential approval change over time, whether we begin from simple
descriptive plots or from ARCH models and cointegration tests on the aggregate series
(Brehm and Gronke, 1995; Ostrom and Smith, 1992). Brehm and Gronke (1995) found
that volatility is predictable: it increases at some points (during election years and after
foreign policy events where the U.S. is perceived to be the aggressor) and decreases at
others (over the course of an administration, during foreign policy events where the U.S.
is perceived to be attacked). Ostrom and Smith do not address volatility per se, but do
refer frequently to an “equilibrium level” of approval, with “errors” from that equilibrium
(perturbations in response to events) that are corrected over time. It may be the case, if
the general hypotheses presented here are true, that the size of these perturbations have
increased over time, and the rapidity of return to equilibrium has decreased, as citizens
have lost their partisan and ideological moorings.
The link between macro and micro approval, to borrow language from Box-Steffensmeier
and DeBoef (1996), is only now being explored. The data portion of this project is daunting; I have accumulated economic and political time series (inflation, unemployment, war
deaths), an event series, and pooled cross sectional surveys for the period 1976–1993. The
eventual goal is to collect a longer set of cross sectional surveys and attaching aggregate
level measure to them, but before taking that step, substantially more work needs to be
completed on determining the particular quantitative models which will be used to analyze
these data. This paper lists a series of alternative models which I hope will stimulate some
reaction and discussion. A modest list includes:
Cointegrated time series and VAR regression: employ the analytic techniques suggested by MacKuen and colleagues, Ostrom, Smith, Box Steffensmeier, and DeBoef.
The time series are not “pooled” in the way that Sayrs (1989) describes. Instead, the
analyst tests for causality across the series. An excellent discussion of these meth16
ods are contained in the volume of Political Analysis in which the Ostrom and Smith
piece appears . In the same volume, however, there is apparent controversy over the
applicability and interpretation of the technology.
Pooled cross sectional time series analyses: instead of testing for integration across
the time series, treat each separate aggregation as an independent observation of the
“average” respondent’s placement on a 0-100 scale of approval. Thus, we are able to
pool time series of whites X blacks X partisans. The problem is that once the sample
is broken down into any moderately high number of categories, the sample size for
each individual time point becomes too small.
Pooled cross sections: another step down in aggregation, this time leaving each of the
individual cross sections untouched, but still estimate pooled models. Markus (1988)
followed this strategy with some success in order to address the “Kramer problem”
in the politics and economics field.
Separate cross sectional analyses: follow a primitive pooling strategy, testing for
stability of coefficients over time. This should allow us to test for change in structural
relationships (e.g. the declining influence of partisanship on approval), but doesn’t
allow a direct modeling of volatility.
This paper has only scratched the surface. I hope this essay, more fairly a work in progress,
serves as a useful resource for ideas, and a target for reactions from others working in this
area. Fire away!
17
Bibliography
Beck, Nathaniel. 1991. “Comparing Dynamic Specifications: The Case of Presidential
Approval.” Political Analysis 3: 51–88.
. 1989. “Estimating Dynamic Models Using Kalman Filtering.” Political Analysis 1: 121–156.
Brace, Paul and Barbara Hinckley. 1992. Follow the Leader: Opinion Polls and Modern
Presidencies. New York: Basic Books.
. 1991. “The Structure of Presidential Approval.” Journal of Politics
53(4): 993–1017.
Brehm, John and Paul Gronke. August, 1995. “History, Heterogeneity, and Presidential Approval.” Working Paper 197 in the Duke University Program in Political
Economy, Durham, NC. Currently under review for publication.
. 1994. Modeling Volatile Approval: A Modified ARCH Approach
to Presidential Approval Ratings. Paper presented at the Annual Meeting of the
American Political Science Association, New York, NY.
Brody, Richard A. 1991. Assessing the President. Stanford, CA: Stanford University
Press.
Engle, R. 1982. “Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of United Kingdom Inflations.” Econometrica 50: 987–1008.
Greene, William H. 1993. Econometric Analysis, 2nd Edition. New York: MacMillan.
Harvey, A. 1990. The Econometric Analysis of Time Series, 2nd Edition. Cambridge,
MA: MIT Press.
Holsti, Ole R. 1992. “Public Opinion and Foreign Policy: Challenges to the AlmondLippmann Consensus. Mershon Series: Research Programs and Debates.” International Studies Quarterly 36: 439–66.
Jackman, Simon. 1995. “Re–thinking Equilibrium Presidential Approval: Markov–
Switching Error–Correction.” Paper presented at the 1995 meeting of the Political
Methodology Section of the American Political Science Association, Bloomington,
IN.
Jentleson, Bruce W. 1992. “The Pretty Prudent Public: Post Post–Vietnam American
Opinion on the Use of Military Force.” International Studies Quarterly 36 (1):
49–74.
Kernell, Samuel. 1993. Going Public (2nd Ed). Washington, D.C.: Congressional
Quarterly Press.
King, Gary. 1993. “The Methodology of Presidential Research.” In Edwards, Kessel,
and Rockman (eds), Researching the Presidency. Pittsburgh, PA: University of
Pittsburgh Press.
. 1989. Unifying Political Methodology. New York: Cambridge University Press.
King, Gary and Lyn Ragsdale. 1988. The Elusive Executive. Washington, DC: Congressional Quarterly Press.
MacKuen, Michael B., Robert S. Erikson, and James A. Stimson. 1995. “Dynamic
Representation.” American Political Science Review, 89(4): 543-565.
. 1989. “Macropartisanship.” American Political Science Review, 83:
1125–1142.
Mueller, John. 1973. War, Presidents, and Public Opinion. New York: John Wiley.
Neustadt, R.E. 1990. Presidential Power: The Politics of Leadership. New York: John
Wiley.
Ostrom, Charles M. and Dennis M. Simon. 1985. “Promise and Performance: A Dynamic Model of Presidential Popularity.” American Political Science Review 79:
334–358.
. 1988. “The President’s Public.” American Journal of Political Science,
32: 1096–1119.
Ostrom, Charles M. and Renée M. Smith. 1992. “Error Correction, Attitude Persistence, and Executive Rewards and Punishments: A Behavioral Theory of Presidential Approval.” In J. Williams (ed.), Political Analysis 4th Ed., Ann Arbor: University of Michigan Press.
Page, Benjamin I. and Robert Y. Shapiro. 1992. The Rational Public: Fifty Years of
Trends in American’s Policy Preferences. Chicago: University of Chicago Press.
Rivers, Douglas, and Nancy L. Rose. 1985. “Passing the president’s program: Public opinion and presidential influence in Congress.” American Journal of Political
Science 29:183–96.
Sayrs, Lois W. 1989. Pooled Time Series Analysis. Newbury Park, CA: Sage Publications.
Simon, Dennis M. and Charles M. Ostrom. 1989. “The Impact of Speeches and Travel
on Presidential Approval.” Public Opinion Quarterly 53(1): 58–82.
Skowronek, Stephen. 1993. The Politicsl Presidents Make. Cambridge, MA: Harvard
University Press.
Stimson, James. 1985. “Regression in Space and Time: A Statistical Essay.” American
Journal of Political Science 29(4): 914–947.
. 1976. “Public Support for American Presidents: A Cyclical Model.”
Public Opinion Quarterly 40: 1–21.
Sullivan, John L. 1973. “Political Correlates of Social, Economic, and Religious Di-
versity in the American States.” Journal of Politics, 35: 70–84.
Zaller, John. 1992. The Nature and Origins of Mass Opinions. New York: Cambridge
University Press.